pouët.net

Soo, seems KeyJ had a point. :)

Frame from my Stargazer capture:


See what's wrong with that frame? :)

A hint, this is one of the intermediate frames the motion estimator dumps out:


Distorted all to hell. In particular, it manages to dump a lot of frames where some of the holes on that menger sponge don't move. So on the resulting motion blurred frame, you see some holes that -should- be fully blurred due to motion [the top and left side of the frame], but aren't.

Sequence:
Real Source Frame -> Filler -> Filler -> Filler -> Filler -> Filler -> Filler -> Real Source Frame


And yes, since someone mentioned it, I am using MVTools and AviSynth mostly for this. I'll share the script once I get something I'm happy with. :)

looks cool, you should put up lifeforce too

micksam7: Yeah, that effect is close to being the worst-case for your pipeline, as there are so many false positive motion patterns here...

The 2424 Debris-video looks absolutely fantastic.

Stargazer at 30 fps with full motion blur

Note how a lot of the distortion effects are near non-existent, also seems to help the demo's motion blur scene too.

Debris uploading..

micksam7: hah, pretty OSSUM. There is some frame-jumping/jitter which is quite annoying at places (the zoomer-part for example), but in other parts with fast motion (the reflective cubes for example), is really really nice.

Aye, there's some jitter with full screen effects and stuff.

Debris at 30 fps

Note that when shaking occurs, things get a bit jittery and blurry as well :)

Haha, Debris was very cool - a crisp demo with a lot of sharp edges and textures = prime candidate for post-release motionblur. :) (..and yes, the hard camera-shake parts look terrible :)

micksam, you're using MVTools 2, right? (with functions like MAnalyse and not MVAnalyse) i once spent a lot of time screwing around with the old version cause I had found an outdated link. just checking :-)

and if so, then i'll shut up :-) (and cheers to keyj)

stargazer cap isnt half bad - i mean it's still an incredibly bad demo (but what do you expect when you let keops work on it..), but a good cap!

This project looks fun. The debris clip looked really good. (Due to terrible wifi at work, i havent seen the other caps yet) Of course, this is yet another nail in the coffin for realtime demos, but i refuse to be a naysayer on such a beautiful day.

Speed: patience, myfriend. In 3-4 years, thanks to postprocessing, even you will enjoy the Cube :D

hyde: it's just keops-banter, i'm a fan allright :)

/me hugs Plek :)

skrebbel: Yes, a modified and carefully tuned 3D recursive search is probably the best motion estimation algorithm to use if you want true motion and don't want to waste too many cycles. Phase Plane Correlation (PPC) might be another good approach, but it's much more computationally intensive and you usually end up doing some 3DRS passes afterwards anyway, because PPC deals with zooms and rotations even worse than 3DRS does ...
But still, there are always situations where the automatic ME algorithms just don't work. And my prediction (though founded more on theoretical knowledge than on actual experience ;) is that it's much worse in demos and games than in natural video or film content.

hrrrmm.. seeing a finished product? might take upto 7 years, but it gets done. somehow ;)

hows progress on your nvision demo btw? :)

skrebbel - Indeed I am :) I need to fiddle with the settings on it a tad more, a few more interesting things I could do with that

hyde - As I kinda mentioned in my first post, this is a silly experiment yes, so I'm not about to add motionblur/random other post processing to my 'real' democaps :D That's too much cheating for me

havoc: ouch :)

the tpb demosystem was acquired by limp ninja a while ago. slowly but surely..

micksam: but it _does_ look good... :)

Interesting. When things aren't going to fast.. it does look more movie'ish.

Just watched Debris 30fps. There sure is blur in there! but a bit too much imho. It's very blurry even in the almost not moving scenes. and those that move a lot you can't see anything.

So yeah, too much blur, or too unstable cams, for my taste.

And when thinking about it, real hi-budget cameras (so what we're used to when watching movies) do not expose the whole 24th of a second but only a fraction of it. So to make it accurate one would have to do some balancing of the frames that get merged together, according to some exposition/shutter speed.

But still, very interresting experiment!

Yeah, I agree.

If you read micksams posts carefully, you'll note that he merges the frames down to double the target framerate, then drops every second frame, effectively resulting in an exposure time of exactly half the frame duration.

But just like depth of field, this is really something you want adjusted properly for each shot individually. I doubt that any one setting will look good on a complete demo.

I suppose you need a sort of adaptive filter, one that decreases blur intensity on static scenes, perhaps analysing motion vector distance changes would be the thing. I wouldn't know where to start though ... maybe Navis would!

ryg: thx for correcting me, I didn't read carefully :)

debris looks fantastic with motion blur (despite the shaky-camera scenes), especially the greetings part.

I experimented with this a while ago. I think the only application of this is to turn a 60fps capture into 24fps with 1/48th film shutter blur to make it more film-like. They're interesting to watch, but nothing should ever be released "final" that way.